The goal is to have an option that avoids symbolic computations/ substitutions for potential speed-ups. Everything should be re-computed at each optimization iteration. Initially, we want to focus on getting it to work without gradients for black-box optimization. If this leads to a speed-up, we could consider extending it using autodiff variables for the snopt optimization.
Why?
Re-computing everything in numeric form might be quicker than substituting the pre-computed symbolic equations with numeric values. This could be the case as the symbolic equations are extremely complex, leading to substitution taking ages.
nepfaff
commented
5 months ago
What?
The goal is to have an option that avoids symbolic computations/ substitutions for potential speed-ups. Everything should be re-computed at each optimization iteration. Initially, we want to focus on getting it to work without gradients for black-box optimization. If this leads to a speed-up, we could consider extending it using autodiff variables for the snopt optimization.
Why?
Re-computing everything in numeric form might be quicker than substituting the pre-computed symbolic equations with numeric values. This could be the case as the symbolic equations are extremely complex, leading to substitution taking ages.